Chl-a model performance summary
Stephanie Clay
2023-02-24
Input data and models
Each chlorophyll-a (chl-a) model uses remote-sensing reflectances (Rrs) as input. The wavebands used in the calculation are dependent on sensor and model. The Satellite Ocean colour and Phytoplankton Ecology group (SOPhyE) at the Bedford Institute of Oceanography uses daily 4km-resolution satellite data from NASA OBPG to calculate chl-a using different models for use in ocean observation and analysis. NASA OBPG reprocesses their datasets every few years as models improve, after which SOPhyE downloads the new datasets and re-optimizes the coefficients used in certain models. Information on reprocessing versions can be found here.
OCI is the standard chlor_a product distributed in files from NASA OBPG, which uses the empirical band ratio model OCx (O’Reilly et al 1998 in combination with a blend of the Hu CI algorithm (Hu et al 2012) for concentrations <= 0.35 mg m3. For the sensors of interest, the OCI product is also referred to by the following combination of acronyms:
- MODIS-Aqua: OC3M, CI
- SeaWiFS: OC4, CI
- VIIRS-SNPP: OC3_VIIRS_SNPP, CI
- OLCI: OC4, CI
POLY4 is a regionally-tuned version of OCx (Clay et al 2019).
GSM_GS is a regionally-tuned version of the semi-analytical GSM model from Maritorena et al (2002). GS refers to the fact that the g coefficients from the original model are spectrally-dependent in this modification (Clay et al 2019).
EOF is a model that employs Principal Component Analysis, currently in use in the Gulf of Saint Lawrence (Laliberte et al 2018).
In situ samples used for validation:
POLY4 and GSMGS models are both trained using in situ HPLC (High
Performance Liquid Chromotography) data. The training set created to
calculate EOF chl-a is composed of satellite matchups to in situ chl-a
derived from Turner fluorescence, as HPLC data is not available for
samples collected in the Gulf of Saint Lawrence.
The oceancolouR
package contains the functions ocx(), gsm(),
and eof_chl() to implement the chl-a models evaluated here.
For eof_chl(), a training set is required for the region of
interest. Using the R2022.0-reprocessed data, the re-optimized POLY4 and
GSMGS are referred to as poly4v2 and gsmgsv2 - e.g. to
use the POLY4 coefficients optimized for R2018.0-reprocessed data, you
would use get_ocx_coefs(sensor, region, alg="poly4"),
replacing sensor with one of modisaqua, seawifs, or
viirssnpp, and region with nwa or nep. To use
POLY4 with R2022.0-reprocessed data, you would retrieve the coefficients
with get_ocx_coefs(sensor, region, alg="poly4v2").
Region of interest
Bounding boxes for each of the regions of interest are defined as follows:
- Northwest Atlantic (NWA): 39-82° N, 42-95° W
- Northeast Pacific (NEP): 46-60° N, 122-140° W
- Gulf of Saint Lawrence (GoSL): 41-53° N, 49-75° W
Matchup restrictions
In situ / satellite matchups used for model training must adhere to the following criteria:
In situ sample must be <= 10 metres from the surface
Satellite pixel and sample location must be within 10 kilometres
The coefficient of variation (standard deviation over the mean) of the OCI chl-a values in the 5x5 pixel box must be <= 0.5
R2018.0 matchups only:
- In the 3x3 pixel box around the matchup satellite pixel, at least 3
pixels must be valid
- In situ sample and satellite pass must occur on the same calendar day
- In the 3x3 pixel box around the matchup satellite pixel, at least 3
pixels must be valid
R2022.0 matchups only:
- In the 5x5 pixel box around the matchup satellite pixel, at least
half the non-land pixels must be valid, down to a minimum of 5 valid
pixels
- Time difference between in situ sampling time and satellite pass
must be within 24 hours and on the same calendar day (see below for
details)
- Solar zenith angle must be <= 75 degrees
- In the 5x5 pixel box around the matchup satellite pixel, at least
half the non-land pixels must be valid, down to a minimum of 5 valid
pixels
Below is a quick comparison of MODIS-Aqua POLY4_v2 satellite chl-a against in situ HPLC chl-a using different restrictions on the difference in time allowed between the in situ sample and satellite pass for a matchup to be used in training and evaluation:
- Sample and pass must be within 12 hours and on the same calendar
day
- Sample and pass must be within 12 hours
- Sample and pass must be within 24 hours and on the same calendar
day
- Sample and pass must be within 24 hours
Using in situ sample/satellite matchups that are within 24 hours of each other on the same calendar day appears to yield the best results, so that is the restriction used in model training.
Disclaimer: The evaluation metrics of the R2018.0 reprocessing here might have slight differences from those presented in Clay et al 2019 due to changes in exact matchup criteria and the order in which the matchups are filtered. The overall message is the same, however, when possible, the latest reprocessing (R2022.0 as of February 2023) should be used.
Chl-a model comparison with the R2018.0-reprocessed files
Northwest Atlantic
Satellite - in situ matchups collected from 1999-2014.
MODIS-Aqua
SeaWiFS
VIIRS-SNPP
Northeast Pacific
Satellite - in situ matchups collected from 2006-2016.
MODIS-Aqua
SeaWiFS
VIIRS-SNPP
Gulf of Saint Lawrence
Satellite - in situ matchups collected from 1997-2019.
MODIS-Aqua
SeaWiFS
VIIRS-SNPP
Chl-a model comparison with the R2022.0-reprocessed files
Northwest Atlantic
Satellite - in situ matchups collected from 2002-2021
MODIS-Aqua
SeaWiFS
VIIRS-SNPP
OLCI
Gulf of Saint Lawrence
Satellite - in situ matchups collected from 2002-2021.
MODIS-Aqua
SeaWiFS
VIIRS-SNPP
